How to Edit Text in an Image: A Tested Workflow
A practical image-text editing guide with actual poster and menu results, 2K dimensions, review steps and the limits of background preservation.
A flattened image is a grid of pixels. A word that was editable in a design app becomes part of the picture when the file is exported. This guide walks through a small, supervised correction using our actual editing pipeline, then shows what the result does and does not establish.
The examples below use artwork we created for controlled tests. They are actual model outputs processed by the editor’s browser compositor. They are not customer submissions, camera photographs or a representative benchmark of every design.
1. Decide whether the image editor is the right tool
If you still have the layered Figma, Canva, PSD or slide file, edit the text object there. That is the best route when the font, kerning and layout must be exact. An image editor is useful when only a PNG, JPG or WebP remains and you have permission to change it.
Start with one short phrase. Rewriting a whole paragraph asks the model to make decisions about line length, spacing and the surrounding artwork. A one-line correction is easier to specify and easier to reject if something goes wrong.
Do not upload confidential material. In this version, image files use public-by-link storage URLs and the configured provider receives the image to process the request. Read our Privacy Policy before uploading anything sensitive.
2. Upload the best source you have
Sign in, then upload a PNG, JPEG or WebP under 10 MB. Use the original export rather than a copy that has been repeatedly compressed or downloaded from a messaging app. Converting a blurry JPG to PNG cannot recover missing letter detail.
For this poster test, the source was 1200 × 800 pixels. The requested operation was the exact phrase JUNE 12 to JUNE 19. The image was owned test artwork with large lettering and a simple illustrated background.
Owned test artwork, before and after. Actual Nano Banana Pro 2K request: 41 seconds observed, 2.1% pixel-difference estimate. One short date on owned flat poster artwork; this does not test dense small print or print readiness.
The request returned a 2528 × 1686 image at the provider’s 2K tier. “2K” is a model resolution setting, not a promise that every aspect ratio produces a 2048-pixel-wide file. Check the dimensions of the file you actually receive.
3. Name the complete phrase
Choose Replace text mode. Type the old wording exactly as it appears, including capitals and punctuation, then type the new wording. If a number appears several times, include an adjacent item name or label so the intended occurrence is less ambiguous.
Our menu test used LATTE 4.50 rather than only 4.50. The replacement was LATTE 5.00. That makes the requested price change specific, although it still does not guarantee that the model will leave other numbers untouched.
Owned test artwork, before and after. Actual Nano Banana Pro 2K request: 48 seconds observed, 3.7% pixel-difference estimate. The changed price is legible; this does not validate allergens, ingredients or a full multi-column menu.
Read every neighboring price and decimal point yourself. The model cannot establish the correct price, ingredient list, allergen information or legal accuracy of a menu. Those facts must come from your own verified source.
4. Separate a generated file from a correct edit
When processing finishes, inspect the complete image. A successful provider task means a file was returned and stored. It does not prove the right characters were drawn or that the background remained suitable.
The editor shows a changed-area estimate and, when possible, a close-up. Its pixel comparison does not read the replacement text. A small changed percentage can still contain a wrong letter. Check spelling at full size, then look at the result at the size where people will actually use it.
If the result changes too much or has a different shape, the verification panel warns you. Do not treat that warning as a successful background match. Keep the original and either try a smaller correction or return to a design tool.
5. Download only after review
PNG avoids adding another lossy compression step. JPG can be smaller, but its white background replaces transparency in the composited export. Both options retain the larger working dimensions when a model result is larger than the source.
Restoring the original background at a larger size involves scaling that source. It does not create photographed detail that the source never captured. For print, compare actual pixels with your printer’s required size and inspect the edge of the edited region.
What our first test set can tell you
We ran 18 live provider requests across replacement, correction, removal, translation and CJK tasks. All 18 returned stored files. The observed median was 39.1 seconds, with a range of 35.2–61.3 seconds. These timings include source upload, provider work, polling and saving the result.
Two results needed additional visual review: a repeating background changed locally and text removal left a faint tone patch. The sample is small, uses deliberately readable artwork and does not establish a production success rate for arbitrary photographs. See the full examples, statistics methodology and machine-readable test record.
Each submitted edit uses one site credit. A technically failed task refunds that credit automatically. A returned image can still be visually unsuitable, so review the actual output before using it.